Automatic lung sound cycle extraction from single and multichannel acoustic recordings

Abstract

A lung sound signal (LSS) consists of a series of inhalations–exhalations or lung sound cycles (LSCs), which provide valuable information about the status of the lungs. Currently, semiautomatic techniques are used to extract LSCs from an LSS. These face limitations in terms of extra cost and effort due to the need of reference signal and inconvenience caused to the patients. Automatic LSC extraction from lung sound (LS) recording can overcome these limitations. In this work, a novel signal processing based method is proposed for extraction of LSCs automatically. At first, the log variance features are calculated from preprocessed LSS, which represent an approximation of the respiratory flow and envelope, but it exhibits spikes. A novel filter-based approach is implemented to smoothen the envelope for better representation of the LSCs’ onset and offset points. The filtered envelopes representing the LS flow are selected through a majority voting technique employing single and multichannel frameworks. The study is conducted on 32 normal and 90 diseased subjects. The mean accuracy (ACC) and onset–offset error (τ) observed for normal category are 94.61% and 0.22 s for both multichannel and single channel frameworks. The same for diseased categories are within 88.96–94.18% and 0.31–0.34 s in multichannel framework and 86.75–92.73% and 0.28–0.34 s in single channel framework. These results are found to be superior when compared with a recently proposed method. The work addresses an important step towards non-invasive computer-aided LS analysis by automated segmentation of LSS without using any additional sensor.